The Origin of Layer Structure Artifacts in ... - ACS Publications

In a recent paper, Yonetani1 reported a strange artifact: when. TIP3P water2 is simulated with group-based truncation and a cutoff of 1.8 nm it forms ...
0 downloads 0 Views 420KB Size
J. Chem. Theory Comput. 2006, 2, 1-11

1

The Origin of Layer Structure Artifacts in Simulations of Liquid Water David van der Spoel* and Paul J. van Maaren Department of Cell and Molecular Biology, Uppsala UniVersity, Husargatan 3, Box 596, SE-751 24 Uppsala, Sweden Received September 9, 2005

Abstract: A recent paper (Yonetani, Chem. Phys. Lett. 2005, 406, 49-53) shows that in computer simulations of TIP3P water (Jorgensen et al. J. Chem. Phys. 1983, 79, 926-935) a strange layer formation can occur when a long cutoff is used. This result is counterintuitive because, in principle, increasing the cutoff should give more accurate results. Here we test this finding for different water models and try to explain why layer formation occurs. In doing so we find that under certain conditions, layer formation coincides with a sharp density increase to 1050 g/L, while simultaneously a pressure of 600 bar develops and water diffusion becomes anisotropic. This leads us to conclude that a group-based cutoff (of at least 1.4 nm) stabilizes an anomalous phase with most water models. In some cases the ordering is strengthened further by periodicity in the simulation cell, but periodicity effects can even be observed with a short cutoff (0.9 nm) and a relatively large box of 4 nm. Water models that have a relatively large quadrupole moment, more in accord with the experimental gas-phase values, in particular TIP4P (Jorgensen et al. J. Chem. Phys. 1983, 79, 926-935), are much less affected by the problem, because the dipole-dipole interaction is quenched at long distance. A comparison of different cutoff treatments, namely truncation, reaction field, particle mesh Ewald (PME), and switch and shift functions, for the simulation of water shows that only PME and shift functions yield realistic dipole-dipole interactions at long distance. The impact for biomolecular simulations is discussed.

1. Introduction Yonetani1

In a recent paper, reported a strange artifact: when TIP3P water2 is simulated with group-based truncation and a cutoff of 1.8 nm it forms liquid layers in which the molecular dipoles are mainly oriented in the same direction.1 The artifact was reproduced in two different MD packages, Amber3 and GROMACS,4-6 implying that it is not due to a software bug. The layers were also found in larger boxes,1 indicating that the behavior is a truncation problem rather than being related to the use of periodic boundary conditions. Interestingly, the layer formation does not occur in conjunction with the particle-mesh Ewald (PME)7,8 method for computing electrostatic interactions. PME is one of the most popular implementations of the Ewald technique, because it * Corresponding author phone: 46-18-4714205; fax: 46-18511755; e-mail: [email protected].

10.1021/ct0502256 CCC: $33.50

is efficient and fast. The electrostatic energy is split into two parts, the short range is computed in direct space and the long range in reciprocal space, using fast Fourier transforms. Other algorithms in practical use for the computation of longrange Coulomb interactions are the particle-particle particlemesh method,9,10 fast multipole methods,11,12 and Lekner summation.13-16 Here we present a series of simulations, using different water models and simulation conditions, in an attempt to explain the phenomenon. For a particular set of parameters we find that TIP3P water transforms to a state with a density of 1050 g/L, a pressure of 600 bar, and a potential energy lowered by 1 kJ/mol compared to the normal -40.8 kJ/mol.17 Similar results are obtained for TIP5P, SPC/E, and SPC water. The simultaneous increase in density and pressure is driven by an increased number of hydrogen bonds and a clearly different, layered, structure; in addition diffusion © 2006 American Chemical Society

Published on Web 11/25/2005

2 J. Chem. Theory Comput., Vol. 2, No. 1, 2006

becomes anisotropic. The combination of these observations leads to the conclusion that this is an artificial phase of liquid water. Complex phenomena such as phase transitions are particularly sensitive to the correctness of the simulation conditions. Slovak and Tanaka have shown18 in simulation studies of melting of ice VII that with a short (0.8655 nm) “smoothly” truncated potential the melting temperatures are very different from those obtained with Ewald summation. In other studies of phase transitions Zangi and Mark19 used a twin-range cutoff with reaction field for the Coulomb interaction, while Matsumoto et al.20 used a shifted potential as first described by Ohmine et al.21 (Appendix B). From a personal communication we learned that Yamada et al.22 used a plain cutoff of 0.9 nm (group-based truncation23,24), whereas Koga et al.25 used a cutoff of 0.875 nm with a switching function. Since there is such a plethora of methods for cutoff treatment, it is very important that authors document their work sufficiently26,27 to allow others to verify it. This work focuses on the treatment of electrostatic interactions, but in principle the accuracy of other simulation algorithms, like integrators and constraint treatment, need to be considered as well. The impact of those algorithms on accuracy fall outside the scope of this paper, however.

2. Methods Molecular dynamics simulations were performed using the TIP3P and TIP4P water models,2 the TIP5P model,28 the SPC model,29 and the SPC/E model.30 Berendsen temperature coupling (298.15 K) and pressure coupling (1 bar) were used.31 The temperature coupling constant τT was 0.1 ps, and the pressure coupling constant τP was 0.2 ps unless otherwise stated. A compressibility of 0.00005 (1/bar) was used. Although we realize that τP is unusually short, we used this value to be compatible with Yonetani;1 in addition, we explicitly tested the influence of τP as described in the results. In all cases a cutoff was used for the Lennard-Jones interactions, but long-range corrections to the energy were applied in the standard way.23 Neighborlists were used and updated every fifth integration time step, which was 2 fs. The water molecules were kept rigid using the SETTLE algorithm.32 Center of mass motion of the simulation box was removed at every time step.33 Five different cutoff schemes were used: 1. a group-based cutoff (simple truncation at the indicated cutoff distance), 2. a reaction field34-36 with rf ) 78.5 (Appendix A), 3. the particle-mesh Ewald algorithm,7,8 4. an atom-based switch function, and 5. an atom-based shift function (Appendix B). For all simulations molecule-based neighbor searching was done, and for both the cutoff and reaction field it should be noted that the cutoff was based on molecules as well, to avoid artifacts due to non-neutral groups. The atom-based switch and shift functions go to zero smoothly; however, it is known that atom based switch functions can cause artifacts when the switching range is too short.24 When using PME the gridspacing was 0.12 nm (fluctuating slightly due to pressure coupling), and fourth-order B-splines were used for charge spreading and force interpolation. Conducting boundary conditions were used for PME.

van der Spoel and van Maaren Table 1. Properties of the Water Molecules Used: Dipole µ, the Components of the Quadrupole Tensor Θ, the Root Mean Square Deviation of the Quadrupole Tensor Elements from the Experimental Values Θyy Θzz RMSD(Θ) Θxx model µ (D) (10-1D nm) (10-1D nm) (10-1D nm) (10-1D nm) expt TIP3P TIP4P TIP5P SPC SPC/E

1.855 2.35 2.18 2.29 2.274 2.351

-2.50 -1.68 -2.09 -1.48 -1.82 -1.88

-0.13 -0.08 -0.11 -0.17 -0.29 -0.30

2.63 1.76 2.20 1.65 2.11 2.19

1.20 0.59 1.42 0.87 0.78

An overview of the simulations performed and the conditions is given in Table 2. All simulations were 2 ns long unless otherwise stated. In total well over 100 simulations were performed, all using the GROMACS software.4-6 All simulations used single-precision arithmetic, except one, which was performed in order to check the effect of precision.

3. Results 3.1. Dipole-Dipole Correlation. Analysis of the simulations focuses on dipole orientation; in particular, we look at the distance dependent Kirkwood factor Gk(r)36 according to Gk(r) )

µi‚µj

∑ r